- articleWhich image formats store straight versus premultiplied alpha, and what does each specification say an agent must do when moving pixels between them?
Alpha conventions differ by format: PNG and WebP store straight (unassociated) colour, OpenEXR stores colour premultiplied by convention; TIFF signals associated vs unassociated per file with ExtraSamples, and AVIF signals premultiplication with a 'prem' item reference (both from producer reading). Never assume one convention across formats; read the format rule or the per-file flag.
- articleWhere does OpenEXR v3.5.2's default scanline deep compositor differ from the tidy/flatten procedure in the documents?
OpenEXR's default deep compositor sorts samples by Z, then ZBack, then input index (only when compositing more than one source) and accumulates them front to back with an early exit at full alpha; it has no split or merge step, so for overlapping volumes it does not implement the documented tidy-then-flatten procedure, and a single-source pixel is assumed already sorted.
- articleDoes OpenEXR v3.5.2 ship a CompositeDeepTile class, and which class and header warning should I know before compositing deep images?
At OpenEXR v3.5.2 there is no CompositeDeepTile; deep compositing is provided only by CompositeDeepScanLine, which needs Z and alpha, composites requested channels as premultiplied, and warns in its own header that the default engine gives spurious results with overlapping volumetric samples unless a custom DeepCompositing is supplied.
- articleIf I move review notes between OpenRV and an OTIO timeline, which survive: markers, drawn annotations, hold/ghost settings?
At Open RV 27db5464 the OTIO reader imports markers (name, range, colour, metadata as JSON) and turns OTIO annotation layers into RVPaint strokes placed on each layer's end-inclusive frame, with hold/ghost as session settings; the writer exports markers only when the importer-created otio_metadata property exists and has no annotation/paint code, so drawn notes are import-only and RV-native markers may not export (inference from the guard). No Marker.comment is carried either way.
- articleHow do hdPrman and Karma name the AOVs/EXR channels they write from USD RenderVars?
At OpenUSD v26.08, hdPrman names each RenderMan AOV after the RenderVar prim name and passes sourceName (prefixed 'lpe:' for LPE sources); the display driver comes from the product file extension (exr/tif/tiff/png, otherwise openexr with a warning; deepRaster + .exr -> deepexr). The render settings prim drives the pass only when the env setting is on or the pass has no AOV bindings, so the same scene can name outputs by different paths. Karma/husk instead derives EXR channel suffixes from the RenderVar data type, with an opt-in lower-case flag.
- articleIn Open RV v4.0.2, does the renderer ever read an RVPaint stroke's softDeleted or uuid property, and are soft-deleted strokes drawn?
Open RV draws an annotation only if its component name is in the frame's order list (then filters by eye, point count and startFrame/duration); no render code reads softDeleted or uuid. Open RV's own writers (annotate mode and the OTIO hook) keep deleted items out of the order list, so a softDeleted=1 stroke would be drawn only if a third-party writer left it in the order list (code inference).
- articleExactly which GTO components and typed properties does Open RV create for a pen stroke, a text box and a shape, and which of them does the renderer actually read?
Annotate mode creates typed properties per object (pen: color, width, brush, points, join/cap/splat, startFrame/duration/mode, plus uuid and softDeleted; text: position, colour, size/scale/rotation, font fields, text, plus uuid and softDeleted; shapes: startFrame, duration, eye and geometry). PaintIPNode, which draws them, reads the drawing properties but never uuid or softDeleted.
- articleIn Kitsu's Zou server, how are shots, tasks, task statuses, review previews and review comments modelled?
Zou v1.0.97 stores shots, assets, sequences, episodes and similar as rows of one polymorphic Entity table (entity_type_id, parent_id); a Task joins an entity, a task type and a status (unique per name, project, type and entity); statuses carry flags such as is_retake and is_feedback_request; comments attach to any object and can change a task's status; a PreviewFile (revision, validation status, JSONB annotations; unique per name, task and revision) is the reviewable artefact, not a publication.
- articleWhere does the OpenEXR library's own deep compositing and deepImageState handling differ from or go beyond the Interpreting Deep Pixels document?
The v3.5.2 default compositor sorts samples by Z then ZBack only when combining more than one source, then composites front to back with over and stops at alpha >= 1; it performs no volume splitting or overlap merging and does not consult deepImageState, so single unsorted images are composited in stored order. The library's SORTED definition uses strict < on ZBack where the document uses <=, and CompositeDeepScanLine requires an alpha channel. Code reading only.
- articleHow are deep images merged, made tidy and flattened to a flat image according to the OpenEXR deep-pixel document, including how alpha and colour are combined?
The document defines merging two deep images as concatenating sample lists; flattening needs a tidy image (sorted and non-overlapping): split partially overlapping samples, merge coincident ones (alpha = 1-(1-a_i)(1-a_j)), sort, then composite front to back with over. Volume opacity follows 1-(1-alpha)^x so splitting preserves total opacity, and the flattened depth is not uniquely defined. Files need not be tidy.
- articleWhat is a deep sample in OpenEXR, and how do I tell a point sample from a volume sample and which channels a deep file must have?
An OpenEXR deep pixel holds any number of samples (all channels share the count), each with a depth; Z is required and ZBack optional (copied from Z if absent); Z >= ZBack is a point sample (zero-thickness surface hit) and Z < ZBack a volume sample over [Z, ZBack); each colour or auxiliary channel needs an associated alpha found through its layer and enclosing layers. The library stores the data; the Interpreting Deep Pixels document defines the meaning.
- articleWhat survives a round trip through OpenRaster for blend modes, groups, adjustment layers and bit depth, per the spec and Krita's implementation?
OpenRaster baseline fixes a closed svg: composite-op list (default svg:src-over) and isolation semantics but leaves pixel storage and blend mathematics to referenced specs. Krita writes blend modes outside that list as krita:<id> (exclusion mapping commented out), saves adjustment layers as a filter element carrying only an applications:krita:<name> type, and reloads them with the filter's default configuration into an rgb8 image, so filter settings and higher bit depth do not round-trip through this code path (code inference).
- articleIn an OpenEXR handed to compositing, how are layers, parts and channel names defined, and what does Blender's writer do when it writes single-part layers versus multipart?
OpenEXR names layered channels L.C (nestable) without any stacking or compositing meaning; multipart parts need unique names (dots allowed) and may differ in compression and data window and mix deep and flat. Blender v5.2.2 writes either single-part channels layerpass.view.C or, in multipart mode, one part named layerpass-view with channels layerpass.C.
- articleHow does UsdRender describe render settings, products and AOVs (RenderVars), and what does it leave undefined (naming, file layout)?
UsdRender (v26.08) describes outputs as RenderSettings -> RenderProduct (raster or deepRaster; productName passed as authored to the driver) -> RenderVar (prim name is the variable name; sourceType raw/primvar/lpe/intrinsic, with intrinsic unimplemented and LPE extensions non-portable). USD sets no universal RenderVar names, and per-product camera/resolution overrides may force separate renders.
- articleDoes libvips resize premultiply alpha, and what pixel-centre offset does it apply?
vips_resize does not premultiply alpha (callers must premultiply first) and applies a 0.5*(1-1/scale) centre-sampling shift except for nearest-neighbour.
- articleWhich pixel-centre convention and alpha handling does OpenImageIO use in resize, resample and warp, and does resize premultiply?
OpenImageIO's resize/resample/warp code uses a half-pixel centre convention, and the transform file does no premultiplication: alpha is filtered like other channels unless the caller premultiplies (the docs provide separate premult/unpremult functions).
- articleEquation by equation, does the OpenEXR repository's TheoryDeepPixels.rst match the 2018-amended PDF, and where does it differ from the 2013 text?
The rst follows the 2018 form at equations 6-7, omits author/date/amendment context, uses 11 Sphinx labels instead of the PDF's numbered equations, and keeps the merge case-4 typo ('alpha_a=1 and alpha=1'), which implementers should read as alpha_b=1.
- articleWhat does Hillman's original paper contain that OpenEXR's current TheoryDeepPixels web page does not show, and are the formulas the same?
The OpenEXR website's TheoryDeepPixels.rst (69b2604f, titled 'Theory of Deep Samples', SPDX BSD-3-Clause) follows the 2018 amendment of Hillman's paper at equations 6-7 (t = T^(1/n)) and keeps the derivation structure, but carries no author, organisation, date or amendment note.
- articleHow does OpenUSD UsdVol represent OpenVDB volumes and their animation over time?
At OpenUSD v26.08 (ee47c679), a Volume binds fields through 'field'-prefixed relationships, an OpenVDBAsset field is animated by authoring filePath time samples per frame (the overview example maps frames 101-104 to separate .vdb files) because '$F' pattern substitution is not supported, fieldDataType must be authored (a missing value is an error) and fieldClass maps to OpenVDB grid classes; the same schema file defines ParticleField prims.
- articleAt the replacement Netflix Studio Partner page, how are VFX plates named, versioned and padded, and when is the plate number versus the plate version incremented?
Netflix's 'VFX Plate Naming Best Practices' page names plates as shot name + plate number + plate version, then frame number and extension (AGM_104_065_010_PL01_v001.[1001-1058].exr), with versions starting with v and typically 3-4 digits, legal characters limited to letters, digits, period, underscore and hyphen, optional typed or word plate codes, and two options for head/tail extensions.
- articleAt the replacement Netflix Studio Partner page, what fields, separators and padding does the VFX shot and version naming recommendation state?
Netflix's replacement 'VFX Shot and Version Naming Recommendations' page (fetched 22 Oct 2026) recommends underscore-separated fields in the order showID, episode, seq, scene, shotID#, task, vendorID, version (example AGM_104_TCC_067_0010_comp_NFX_v001), 3-4 digit shot IDs usually incremented by 10, versions beginning with v, and consistent field inclusion across vendors and shots within a project.
- articleWhat exactly is stored in a Cryptomatte EXR (hashing, channel layout, manifest) and how should an agent read or write it safely?
Cryptomatte 1.2.0 (repository commit 968d5e4b) suggests 6 ranks in three numbered EXR layers (00-02) with preview channels deprecated, hashes UTF-8 names with MurmurHash3_x86_32, converts to float32 by bit-copy with the exponent clamped to [1:254], supports little-endian only, and gives a test vector ('torus' -> 4053562365, -1.54943624832e+30); OpenEXR's Deep IDs document notes Cryptomatte's fixed per-pixel ID count, FLOAT ID channels that must not be colour-managed or rewritten as HALF, and an ID space just under 2^32.
- articleDoes any openly readable on-set VFX data template record who captured each item (operator/person) and whether it was checked or unverified, and if not what is the nearest documented substitute?
Negative result with partial analogues. In the VES Camera Report spec the only person field is slate-level 'Wrangler' (person entering data); the VES On-Set VFX Data Guide (v1.1.0 file) assigns a Creator role per data item (role names, not persons). Across the pinned VES texts read there is no verified/unverified or checked field; the nearest items are the 'Under Protest' VFX Work flag, the guide's Clips Report 'data confirmation' and the timestamped virtual-production scene snapshot. A synthesis proposal adds 'Tk Captured By' and 'Tk Verification Status' columns, which the spec's own extension rule and sample reader would carry (untested).
- articleI need a free, ready-made on-set VFX camera report: which fields does the VES form have, what person field exists, and what does it let the wrangler tick as captured?
The VES Camera Report (VES Technology Committee, camerareports.org) is a free Filemaker database, blank paper PDF and CSV interchange spec whose slate/take fields cover lens (focal length, focus distance, lens height, serial, stop, shutter, tilt, dutch), unit, camera and a 'Set Media' checklist (Balls, Clean Plate, Cube, HDRI, Mocap, Photos, Survey Diagram, Tile Set, Virtual Set, Witness Cam). Its only person field is slate-level 'Wrangler' ('person entering the data'); no verification/checked/status field appears in the spec, proposal, changelog, developer guide or blank form texts read. LIDAR is not a baseline Set Media value, though the spec says other values may be defined.
- articleWhat do the RVM, MatAnyone and MatAnyone 2 papers report on video-matting benchmarks (MAD, MSE, Grad, Conn, dtSSD), how do synthetic and real-video numbers differ, what limits do they state, and what are the code and weights licences?
RVM (arXiv v1, GPL-3.0 code) is auxiliary-free human matting and states it fails when people are in the background. MatAnyone (v2) and MatAnyone 2 (v2) are mask-guided: they take a first-frame mask, MatAnyone states it has the best MAD and dtSSD on every synthetic benchmark block (not every MSE or Grad), and it says the older VideoMatte240K ground truth lacks fine detail such as hair. Numbers do not transfer between benchmarks: MatAnyone reports core-region MAD 0.14 on a real-video benchmark with segmentation-mask ground truth, while MatAnyone 2 reports MatAnyone at MAD 5.76 on 19 real videos with manually annotated alpha every 10 frames. Even the newest model has 15.2% of boundary-band pixels flagged unreliable by the paper's own learned evaluator, whose checkpoint is not released at the pinned README. MatAnyone and MatAnyone 2 code carries the non-commercial S-Lab licence.
- articleWhat J&F accuracy does the SAM 2 paper report for video object segmentation, where does it say SAM 2 fails, which checkpoint does the public README actually describe, and what licence covers each component?
Paper v2 reports SAM 2 (Hiera-B+) J&F of 76.6 on MOSE val, 90.2 on DAVIS 2017 val, 78.0 on LVOS val and about 77 on SA-V val/test, ahead of Cutie-base+ (71.7, 88.1, -, 61.3/62.8) when prompted with the first-frame ground-truth mask. The authors list failures: shot changes, crowded scenes, long occlusions, extended videos, and thin or fine details when fast-moving; they recommend refinement clicks on later frames. The public README's SAM 2.1 per-checkpoint numbers (large: SA-V test 79.5, MOSE val 74.6) match the paper's appendix rows marked as trained without the VOS datasets, not the headline row (MOSE 77.9), which is a documented difference an agent must not collapse. Licences: checkpoints, demo and training code Apache-2.0 (README), SA-V dataset CC BY 4.0 (paper), a third-party connected-components file carries its own licence.
- articleWhat do the VES on-set data guide and the tracking tools say about survey measurements and witness cameras for camera tracking, and what do they leave unspecified?
The VES guide (title page 1.1.0, 2026-05-12) lists witness cameras, lidar, photogrammetry, lens encoders and reference-camera timecode as items to capture, each with purpose and consumers, but gives no quantities (no camera counts, marker sizes or tolerances) in the sections read. Tool manuals say how survey data is consumed: Nuke ties sequences to a known 3D world via survey points; Blender sets scale from a distance in scene units; SynthEyes keeps "survey shots" (stills) distinct from set surveys. The term "witness camera" means extra video angles in the VES guide and unordered stills in the SynthEyes manual.
- articleHow should a lens grid be shot on set so a tracker or compositing tool can use it, and what goes wrong with a bad grid?
A 2026-05-01 practitioner page gives a field checklist: rigid, flat, level grid; a long dolly track; the grid filling the whole frame; sharp focus; even light at about F11 for spherical lenses; and a slate with body, lens, actual focal length and focus distance. It quotes a supervisor against mixed-resolution grids and tape over intersections. Nuke’s grid-analysis page states it should be used when a grid exists and gives a recovery step when detection under-covers the image; Blender’s manual says it has no calibration tool and points to OpenCV.
- articleWhat does HOOD, an open learned cloth simulator trained without a ground-truth solver, report about quality, speed and limits compared with MeshGraphNets-style and physics-based cloth?
HOOD (CVPR 2023, arXiv v3) combines graph networks with hierarchical message passing and trains self-supervised with a physics-based loss instead of solver-generated targets. Its human study (30 participants, pairwise realism) favours HOOD over two learned baselines and is about equal to the ArcSim physical simulator on the sequences where ArcSim could run; HOOD runs at about 7-14 fps for a 12K-vertex dress ablation (not a comparison with ArcSim time). Stated limits: no garment-garment interaction (self-penetration), failure when body motion exceeds training velocities, and wrong correspondences under severe body self-intersection. Code is MIT-licensed per the LICENSE text; data and SMPL models are fetched separately.
- articleWhat do studio and vendor first-person accounts say about the real speed and accuracy gains of ML deformers in production rigs, and where do they fall short?
DreamWorks and Berkeley (2018) report 5x-10x faster deformation evaluation on production rigs. Golaem (DigiPro 2024) tried to reproduce this in a commercial crowd product and reports only x1.2-x2 on production assets because the paper's rigs were optimised for the model and the rest of rig evaluation dominates; its main disappointment was no significant speed gain. It reports Euler-angle and facial limits of the first method, ships the ML deformer as a crowd-package feature for close-up facial deformation, and notes crowd artists prefer overfitting because crowd sims blend clips rather than generate unseen animation. The paper's own Table 1 fps values imply 1.8x-2.5x, which does not match its text range.
- articleUnder the Move One App EULA, who owns the mocap Output, what may I do with it commercially, and what may I not do with it (AI training, competing products)?
At the Move One App EULA page fetched 2026-10-22 (header Last-Modified 2026-10-22T01:02:33Z), Move AI owns the IP in the Motion Data and Output (7.1.2) and licenses the Output back to the user worldwide, royalty-free, transferable and sublicensable (7.3), but the licence is void if Output is put to a defined Prohibited Use (7.5): building a competing product, training any AI System, or using it as an input to an AI System other than to create an original video work. The user in turn grants Move a perpetual, irrevocable licence to uploaded video for providing services and for developing and training its products (7.2). This is clause reading only, not legal advice or clearance.
- articleWhere do the lens-distortion models of matchmove and compositing tools (Nuke, SynthEyes, OpenCV, Blender/libmv, COLMAP, 3DE, OpenTrackIO) disagree, and what parameters must travel with a shot for the distortion to be reproduced?
Reading pinned vendor docs and source, the models disagree in at least five ways: direction (Blender division applies x/(1+k r^2) from ideal normalised to image; Nuke Classic and COLMAP SIMPLE_DIVISION use the same form from distorted to ideal), the meaning of k4 (OpenCV: a rational denominator term; Blender Brown: an r^8 polynomial term), the roles of p1 and p2 (Blender Brown is swapped relative to OpenCV, COLMAP and Blender Polynomial, from code reading), the normalisation (OpenCV by z with focal in pixels; Nuke width/height/diagonal/maximum; Blender Nuke model by half the larger image side; OpenTrackIO offsets in mm), and pixel origin conventions. For the shot to be reproducible the record must carry model name, direction, coefficient order and units, normalisation, centre/offset, sensor size and resolution, focal length and the focus/zoom state; an STMap or projection screen is the model-free alternative with storage and precision costs. 3DEqualizer model equations were not read...
- articleDoes the Alembic project publish sample .abc files, and what terms cover them?
The Alembic repository does not publish a curated sample-scene set. Its 46 .abc files are test inputs: mostly fuzzer and malformed-input regressions and security proof-of-concept files, plus two small renderer test files (prman cube.abc and xforms.abc). The only licence file is a BSD 3-clause text addressed to source and binary forms, which does not mention the .abc data, so reuse terms for those files are unclear.
- articleWhen lens metadata reports "focal length", which focal length is it, and what reference point do entrance-pupil and focus values use?
Cooke distinguishes nominal, pinhole and effective focal length and warns that mixing them causes matchmove problems; positions are referenced to a nominal image plane 52 mm behind the PL flange; pinhole focal length is available as /i lens metadata on supporting lenses and otherwise must be measured by grid calibration or solved in a match-move. Record which focal length a metadata field carries.
- articleWhich public OpenUSD test assets exist (usd-wg/assets, Pixar Kitchen Set), and which of them state terms that would let L10 use them?
The usd-wg/assets repository (3b75c2da) declares licences per asset in each readme (typically CC0 or highly permissive) with the repository licence as fallback, while Pixar's Kitchen Set is offered only under a temporary, revocable licence for personal, non-commercial testing of USD.
- articleWhat must a custom character satisfy to be used in Flow Studio, which terms govern its outputs, and how have the documented capabilities changed over recent dated releases?
The character requirements page (stamped Oct 19, 2026) sets file, polygon, hair-strand and DCC-version limits, T-pose and add-on validation requirements. The release notes show the platform broadening in 2026 from live-action replacement into a 3D Editor and a Canvas that hosts third-party generative video/audio models, so "Flow Studio" now covers more than the original tool. Output ownership is not stated on the pages read: the only terms linked from the help site are Autodesk's General Terms (Internet Archive snapshot 2026-10-16), which say users keep ownership of what they submit, define "Output", and disclaim responsibility for it; product-specific Special Terms were not read, so Flow Studio output rights remain unresolved.
- articleHow does Physics-IQ test physical understanding in video models, which models and versions were scored, what were the results, and what do the authors say they do and do not show?
Physics-IQ (Motamed et al., Google DeepMind) gives a model the first 3 seconds (or the last frame) of a real, static-camera physics event and asks it to continue for 5 seconds, then compares to the real continuation with four motion-based metrics, combined and normalised so two real takes of the same scene score 100. At v3 the best of eight configurations (VideoPoet, multiframe) scores 29.5 and Sora (image-to-video) 10.0; visual realism judged by Gemini 1.5 Pro was not significantly correlated with physics score (r = -0.46, p = .249). The authors call the metrics proxies, say they are on the conservative side (they penalise camera moves and shot cuts), and say results do not show that future models cannot learn physics.
- articleFor a newcomer: which of GIMP, Krita and Inkscape handle more than 8 bits per channel and colour-managed (ICC) work, and which version should I assume?
As observed on the project listings on 2026-10-21/22, GIMP 3.2.6 (released 2026-09-10) and Krita 5.3.4 / 6.0.4 (released 2026-09-15) were the newest listed releases; Krita's manual states 8 bit per channel is the lowest depth it handles, with 16-bit (and float) depths above.
- articleIn Blender's GPU compositor what does texture_load do for out-of-range reads, and what does that mean for the Keying node's edge test at image borders?
Blender's GPU compositor texture_load(sampler2D, int2) clamps to the edge texel (replicate), only the fallback overload returns a chosen value and texture_load_unbound does not guard; the Keying tweak-matte shader uses the clamping form for every read, so border windows see replicated border values (static reading at f46990f5).
- articleWhat do public docs say a generative background tool for LED volumes (Cuebric) outputs, and what is the .2p5d file format that Disguise uses to receive those layered 2.5D backgrounds?
Cuebric publishes no technical documentation that curl can read: cuebric.com is an Angular shell with marketing text and a four-URL sitemap (home, app login, privacy, terms). The usable public record is on the receiving side: Disguise documents a Cuebric integration (segment and dimensionise a 2D image, export a depth package) and a .2p5d container (renamed zip with data.json and per-plate image/video, optional depth map or OBJ mesh, format version 1.0.0, inverse-normalised depth with an explicit conversion formula). Cuebric capabilities such as input resolution, supported generators, accuracy of segmentation/depth and output terms are not documented in anything read; the "ready to shoot" wording is a vendor-partner claim.
- articleCan I get a free-viewpoint 4D scene from one ordinary (phone) video, and how do I tell whether a monocular dynamic-scene result is real?
The primary sources read say: not reliably. The DyCheck paper (arXiv 2210.13445v1) shows that common monocular benchmarks effectively leak multi-view signal through teleporting cameras or quasi-static scenes, and that with realistic camera motion and complex motion published NeRF-era methods lose roughly 1-2 dB (masked PSNR, no multi-view cues) and 4-5 dB (complex motion) per the abstract. The 4D-GS paper's own appendix reports overfitting to training views on monocular iPhone-style input and failure on large motion, and Dynamic 3D Gaussians states it needs a multi-camera rig. A reader can check a monocular claim by looking at how the camera moves relative to the scene (EMF), whether metrics are co-visibility masked, and whether sequences are not quasi-static. DyCheck evaluated NeRF-era models (T-NeRF, NSFF, Nerfies, HyperNeRF), not Gaussian methods; no read source reports a Gaussian method on its masked metrics.
- articleWhat does it mean to reconstruct a dynamic (4D) scene with 4D Gaussian splatting or a dynamic NeRF, what do I have to capture, and what do I actually get back?
A 4D scene method fits a per-scene model to video so that images can be rendered for a chosen camera and a chosen time. In the papers read, 4D Gaussian splatting keeps one set of 3D Gaussians plus a learned deformation (or native 4D primitives) queried by time; dynamic NeRFs (D-NeRF family) do the same with a canonical radiance field plus a deformation network. What you must capture depends on the method: Dynamic 3D Gaussians needs a synchronised multi-camera rig (27 training cameras in its experiments); 4D-GS and Real-time 4DGS are benchmarked both on 15-20 camera rigs (Neu3D) and on one- or two-camera and synthetic monocular sets; D-NeRF assumes one camera and one image per time instant. The output is a trained model that renders novel views at times inside the captured span (plus optional per-timestamp Gaussian exports), not a rigged or topology-stable mesh. Every number here is author-reported and nothing was run.
- articleBefore an agent hands an AI-generated roto/matte to compositing, what should it check, which of those checks do the sources actually support, and which have no source-given test?
Synthesised from the pinned sources: record tool and model version; decide mask versus alpha; scrub frames at the failure scenarios the authors name (shot changes, crowds, long occlusions, thin fast structure, similar objects in background); check core holes and background specks; compare the edge band with an outline view; judge temporal chatter on a static subject; check the delivery resolution against the model input; bake the matte to pixels; and record licence status. The sources support the failure list and the controls but give no pass/fail threshold; the one ground-truth-free quality metric found (MatAnyone 2 unreliable-pixel rates) depends on an evaluator checkpoint that is not released at the pinned README. Every checklist item below is the lead's synthesis from vendor and paper statements, not a published standard.
- articleWhat failure modes (identity drift, geometry violations, flicker) do sources report when photoreal video is generated from engine renders or structural controls?
Reported failure modes come from papers and vendor caveats, not studio accounts. RealMaster (research, GTA-V) reports identity drift and colour shifts in an Aleph baseline, flicker and identity fluctuation from multiple image-edit anchors, face-structure loss under depth conditioning, and temporal artifacts under fast motion. NVIDIA reports hallucinated or misoriented vehicles for Transfer1 AV variants and error accumulation over long clips, and Transfer1 papers show geometry errors for single controls (segmentation-only headings). Vendor docs warn that high control weight can cost quality and that image-sourced subjects may be inconsistent.
- articleHow should I name a VFX shot and its versions so a studio or vendor will accept the delivery?
Netflix requires shot names showID_episode_scene_VFXID (series; episode 102 = season 1 episode 2) or showID_sequence_scene_VFXID (film), and a CSV Version Name equal to the full movie file name or the EXR base name without frame numbers. A community convention (ComfyUI README after Victor Perez) uses AAA_AAA_####_aaaa_aaa_v###.####.aaa with shots numbered in tens and FINAL for finals. The two differ; studios' own specs govern deliveries.
- sourceDepthCrafter (CVPR 2025 Highlight)
- sourceProPainter (video inpainting, ICCV 2023)
- sourceMatAnyone 2 (CVPR 2026 Highlight)
- sourceMatAnyone (CVPR 2025)
- sourceRobust Video Matting (RVM)